src/pdf_parser.py

# pdf_parser.py

import fitz  # PyMuPDF
import os
import torch
import pickle
import numpy as np
from PIL import Image
import cv2
import pytesseract
from transformers import CLIPProcessor, CLIPModel
import faiss

# Load OpenCLIP model (CPU)
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
model.eval().to("cpu")

def parse_pdf(pdf_path):
    doc = fitz.open(pdf_path)
    metadata = []
    all_embeddings = []

    for i, page in enumerate(doc):
        # Step 1: Convert to image
        pix = page.get_pixmap(dpi=150)
        image_path = f"page_{i+1}.png"
        with open(image_path, "wb") as f:
            f.write(pix.tobytes())

        # Step 2: OCR
        image = Image.open(image_path).convert("RGB")
        ocr_text = pytesseract.image_to_string(image)

        # Step 3: Thumbnail
        thumbnail_path = f"thumb_page_{i+1}.jpg"
        cv_img = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
        thumb = cv2.resize(cv_img, (250, 350))
        cv2.imwrite(thumbnail_path, thumb)

        # Step 4: Vision Embedding
        inputs = processor(images=image, return_tensors="pt").to("cpu")
        with torch.no_grad():
            embedding = model.get_image_features(**inputs)[0].cpu().numpy()
        embedding /= np.linalg.norm(embedding)

        all_embeddings.append(embedding)
        metadata.append({
            "page": i + 1,
            "thumbnail": thumbnail_path,
            "ocr": ocr_text.strip()
        })

    # Step 5: Save FAISS index
    dim = all_embeddings[0].shape[0]
    index = faiss.IndexFlatL2(dim)
    index.add(np.array(all_embeddings).astype("float32"))
    faiss.write_index(index, "vision_index.faiss")

    # Step 6: Save metadata
    with open("metadata.pkl", "wb") as f:
        pickle.dump(metadata, f)

    print(f"✅ Processed {len(doc)} pages.")